Volantis raises $88M to route light around the memory bottleneck Europe is building

On 2 October, The Next Web reported that San Francisco startup Volantis closed an $88M Series A for a photonic memory layer aimed at AI inference. The round was led by Lachy Groom and Abstract Ventures, with John Doerr, VXI Capital, Triatomic and Susa Ventures participating, as TechFundingNews first reported. First customer deliveries are planned for 2027.

The timing is not incidental. Three days earlier, Micron told investors that memory supply will stay tight through 2028, and that about 75% of its 2027 output is already spoken for.

What Volantis is actually claiming

The company is building a photonic link between AI compute and memory, using gallium arsenide lasers rather than indium phosphide. According to The Next Web, the design deliberately routes around indium phosphide because China holds roughly 70% of that material's supply and wafer prices rose about 250% by June, a constraint IQE's chief executive has called a key risk for the semiconductor industry. Volantis says its A-1 system targets models above 20 trillion parameters at up to 10,000 tokens a second per user. Those are design goals rather than measurements from a running machine, and the company has not published benchmark results. The founders are Tapa Ghosh, a Thiel fellow and former Y Combinator founder, and Roy Meade, who ran Micron's high-bandwidth memory programme and was a vice president at Ayar Labs.

Europe is moving in the opposite direction on materials. The Netherlands is building a EUR 153M indium phosphide pilot line at Eindhoven that moves to 6-inch wafers and becomes operational in 2027, the same year Volantis plans to ship. PhotonDelta has opened a Silicon Valley hub to connect the two ecosystems. PIXEurope holds EUR 380M across 20 members in eleven countries, coordinated from Barcelona, with around EUR 133M going to the Netherlands.

The two funding tracks are not buying the same thing. Europe's money is public and industrial, aimed at a factory. Volantis is private capital betting that the material Europe is industrialising is one to design around. Lightmatter has raised $400M, Ayar Labs $500M and Lumilens over $700M, according to The Next Web. Switzerland's Kandou raised $225M in March to argue that copper signalling can outlast the optical transition, because copper is already in every data centre.

Micron's numbers set the backdrop

The demand side of that bet is now unusually well documented. Micron reported record results on Wednesday 30 September for the quarter ending 3 September: revenue of $54.23bn, up nearly fivefold year on year, and net income of $37.7bn against $3.2bn a year earlier. Gross margin hit 86.8%, compared with 44.7%. Analysts polled by LSEG had expected $51.07bn for the quarter, CNBC reported.

Chief executive Sanjay Mehrotra was blunt on the earnings call. "We do not have line of sight to when supply and demand will return to balance," he told investors, according to The Next Web. He said demand will exceed supply in both 2027 and 2028, and that the market will be tighter in those years than in 2026. Most current sales talks, he said, are about 2028.

Ars Technica reported on 1 October that Mehrotra's comments refer to business-to-business sales of high-bandwidth memory for AI and DRAM for servers, since Micron no longer sells consumer RAM. But the implications run downstream, because manufacturing capacity is prioritising AI and servers. "Overall, supply-demand environment is only getting tighter," Mehrotra said, per a transcript from Seeking Alpha cited by Ars Technica.

"Even after they are built, even after first wafer output, production ramps up only gradually in the clean rooms," Mehrotra said, per Ars Technica. "That's just the nature of what it takes to bring up production."

Micron plans new clean rooms in 2028 and about $25bn of capital spending in the first half of its new financial year. Mehrotra said the ramp will not quickly ease supply. He also pointed to HBM moving from 3E to a greater mix of 4 and 4E, and to node transitions delivering less productivity gain per wafer.

Where the margin is coming from

Tom's Hardware reported on 2 October that Micron's Mobile and Client unit, its consumer business, posted an 88% operating margin and was the only unit that shipped less memory in the quarter. Revenue there rose 14% quarter over quarter. Core Data Center gross margin went from 25% to 85%, Automotive and Embedded from 20% to 79%, and Cloud Memory from 48% to 76%. The gap between gross and operating margin in the consumer unit is just 2%, which Tom's Hardware notes means the cost of running that business is very low relative to the revenue it generates.

That is price, not volume. The clearest consumer proof comes from Raspberry Pi, which has raised the price of its 2GB Pi 4 and Pi 5 by $12.50, citing memory costs. The 2GB Pi 4, once a $35 computer, now costs $67.50. The company says the cheap memory it bought in 2025 is largely used up. Samsung has added $100 to most of its Galaxy S26 range.

The capacity arithmetic behind those increases is straightforward. Samsung executive Kim Taewoo said this week that HBM will take almost 30% of the industry's DRAM wafer capacity in 2027, up from 20% this year, Reuters reported. Ars Technica carried the same figure.

New capacity, different geography

China's ChangXin Memory Technologies plans to invest 34.9 billion yuan, about US$5.2bn, in expansion, according to a Monday stock filing reported by the South China Morning Post on 29 September. The filing allocates 24.1 billion yuan to a new technology R&D project and 10.8 billion yuan to the second phase of a wafer testing base. Nearly 22.4 billion yuan, about 93% of the R&D budget, goes to equipment purchases. A source told the SCMP that CXMT will favour domestic suppliers, with an "aggressive" target for local chipmaking equipment.

That expansion does not resolve the near-term crunch, and it is occurring alongside litigation. Tom's Hardware reported on 1 October that Micron filed a lawsuit this month against YMTC, accusing the Chinese memory maker of poaching its engineers, obtaining its NAND know-how, and then using patents granted to those former employees to accuse Micron of infringement in multiple jurisdictions. Micron alleges the employees gave personal or family reasons for resigning and did not update their LinkedIn profiles to reflect new roles at YMTC. YMTC has not commented in the coverage available.

Elsewhere in the stack, the shortage is reshaping what gets built. SemiEngineering reported on 1 October that AI-defined vehicles are pushing compute, memory and validation limits, with automakers split between large central processors and distributed specialised compute, and raw camera and radar data putting new demands on memory and interfaces. A separate SemiEngineering piece the same day examined non-volatile memory requirements for humanoid robots, citing Goldman Sachs analyst estimates of more than 6 million annual humanoid shipments by 2035, up from fewer than 15,000 units in 2025. Telemetry logging at sampling rates above 10 kHz, that piece notes, can exceed the endurance limits of conventional non-volatile memory.

Research on agent memory is pointing the same way. A paper submitted to arXiv on 9 July, Heavy-Tailed Memory Traces in Long-Horizon Language Agents, argues that agent memory concentrates on a small core while rare states sit in a long tail where prediction errors accumulate. The authors report that their Core-Tail World Model reduces prompt tokens by 5.9% and lowers bottom-half tail prediction error by 13.6% versus a graph-memory baseline, with a 24.48% token reduction on LongMemEval at aggregate accuracy parity. A second paper, MemFit, submitted on 1 October, stores each conversational turn verbatim in an append-only store with LLM-free insertion and reports state-of-the-art results on LoCoMo, MemGallery and LongMemEval-S while cutting memory construction cost several-fold.

None of this substitutes for wafers. Micron's forecast for the current quarter is revenue of $61.5bn, give or take $1.5bn, against analyst expectations of $57bn, according to LSEG. DRAM revenue rose 343% year on year to a record $39.8bn, 73% of the total, and NAND flash rose 526% to $14.1bn. Mehrotra also said Micron raised pay for every employee in fiscal 2026, after hundreds of staff in Taiwan had threatened to strike over pay, CNBC noted.

The two stories, photonics and memory pricing, meet at the same constraint. Volantis is selling an architecture that assumes the bottleneck persists. Micron is telling investors it will persist through 2028. Only one of those positions has a shipped product.